DecisionFlow for SMEs: A Lightweight Visual Framework for Multi-Task Joint Prediction and Anomaly Detection
Ruolin Qi
Abstract
The digital transformation of industry has placed immense pressure on small and medium-sized enterprises (SMEs), which often lack the resources and technical infrastructure to adopt complex AI pipelines and data visualization systems. This creates a pressing challenge: how can SMEs leverage real-time business intelligence without heavy computational or financial burdens. We propose the DecisionFlow framework to generate interactive KPI dashboards on the fly using Vega-Lite declarative syntax through a low-latency visualization engine. To run at the edge, we distilled the long sequence of self-attention into linear Performer blocks, using cross-prediction step weight sharing and 8-bit quantization, so that the model only contained 4.3 M parameters and could be deployed on a common CPU or browser WebAssembly environment. Secondly, the model couples continuous probability prediction with rare event detection with a joint uncertainty loss function, as well as a cross-layer visual feedback closed-loop in which the user interacts with real-time update of the attention mask, so as to support online incremental learning. On Online Retail II (UCI), a dataset for SMEs, DecisionFlow reduced inference latency by 68%, demand forecasting MAE by 17%, and anomaly detection F1 to 0.91.